path: root/scripts/debbisect

#!/usr/bin/env python3
#
# Copyright 2020 Johannes Schauer Marin Rodrigues <josch@debian.org>
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.

# snapshot.d.o API feature requests:
#
# Currently, the API does not allow to list all dates for which a snapshot
# was made. This would be useful to allow skipping snapshots. Currently we
# blindly bisect but without knowing which date on snapshot.d.o a given
# timestamp resolves to, we cannot mark it as untestable (see feature request
# above) and without a list of testable timestamps we cannot reliably test
# a timestamp before and after the one to skip.
# See also: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=969603
#
# It would be useful to know when a package version was first seen in a given
# suite. Without this knowledge we cannot reliably pick the snapshot timestamps
# at which we want to test a given suite. For example, a package version might
# appear in experimental long before it appears in unstable or any other suite
# that is to be tested. Thus, the first_seen attribute of the snapshot API is
# not very useful to determine a limited set of timestamps to test.
# See also: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=806329

# wishlist features
# -----------------
#
#  - restrict set of tested snapshot timestamps to those where a given package
#    set actually changed (needs a resolution to #806329)
#
#  - allow marking snapshot timestamps as skippable, for example via script
#    exit code (needs resolution to #969603)
#
#  - add convenience function which builds a given source package and installs
#    its build dependencies automatically

# complains about log_message cannot be fixed because the original function
# names one of its arguments "format" which is also forbidden...
# pylint: disable=W0221
#
# pylint complains about too many branches but the code would not become more
# readable by spreading it out across more functions
# pylint: disable=R0912
#
# allow more than 1000 lines in this file
# pylint: disable=C0302
#
# TODO: Adress invalid names
# pylint: disable=invalid-name

import argparse
import atexit
import collections
import email.utils
import http.client
import http.server
import io
import logging
import lzma
import math
import os
import pty
import re
import select
import shutil
import socketserver
import subprocess
import sys
import tempfile
import threading
import urllib.error
import urllib.request
from datetime import date, datetime, time, timedelta, timezone
from functools import partial
from http import HTTPStatus
from time import sleep

import debian
import debian.deb822
import requests

HAVE_DATEUTIL = True
try:
    import dateutil.parser
except ImportError:
    HAVE_DATEUTIL = False

HAVE_PARSEDATETIME = True
try:
    import parsedatetime
except ImportError:
    HAVE_PARSEDATETIME = False

DINSTALLRATE = 21600


def format_timestamp(timestamp):
    return timestamp.strftime("%Y%m%dT%H%M%SZ")


# We utilize the fact that the HTTP interface of snapshot.d.o responds with a
# HTTP 301 redirect to the corresponding timestamp.
# It would be better if there as an officially documented API function:
# http://bugs.debian.org/969605
def sanitize_timestamp(timestamp):
    conn = http.client.HTTPConnection("snapshot.debian.org")
    conn.request(
        "HEAD", "/archive/debian/" + timestamp.strftime("%Y%m%dT%H%M%SZ") + "/"
    )
    res = conn.getresponse()
    if res.status == 200:
        return timestamp
    assert res.status in (301, 302), res.status  # moved permanently or temporarily
    prefix = "http://snapshot.debian.org/archive/debian/"
    location = res.headers["Location"]
    assert location.startswith(prefix)
    # flake8 wrongly insists that there must be no whitespace before colon
    # See https://github.com/PyCQA/pycodestyle/issues/373
    location = location[len(prefix) :]  # noqa: E203
    return datetime.strptime(location, "%Y%m%dT%H%M%S%z/")


# we use a http proxy for two reasons
#  1. it allows us to cache package data locally which is useful even for
#     single runs because temporally close snapshot timestamps share packages
#     and thus we reduce the load on snapshot.d.o which is also useful because
#  2. snapshot.d.o requires manual bandwidth throttling or else it will cut
#     our TCP connection. Instead of using Acquire::http::Dl-Limit as an apt
#     option we use a proxy to only throttle on the initial download and then
#     serve the data with full speed once we have it locally
class Proxy(http.server.SimpleHTTPRequestHandler):
    def do_GET(self):
        # check validity and extract the timestamp
        try:
            c1, c2, c3, timestamp, _ = self.path.split("/", 4)
        except ValueError:
            logging.error("don't know how to handle this request: %s", self.path)
            self.send_error(HTTPStatus.BAD_REQUEST, f"Bad request path ({self.path})")
            return
        if ["", "archive", "debian"] != [c1, c2, c3]:
            logging.error("don't know how to handle this request: %s", self.path)
            self.send_error(HTTPStatus.BAD_REQUEST, f"Bad request path ({self.path})")
            return
        # make sure the pool directory is symlinked to the global pool
        linkname = os.path.join(self.directory, c2, c3, timestamp, "pool")
        if not os.path.exists(linkname):
            os.makedirs(os.path.join(self.directory, c2, c3, timestamp), exist_ok=True)
            os.symlink("../../../pool", linkname)
        path = os.path.abspath(self.translate_path(self.path))
        if not os.path.exists(path):
            self._download_new(path)
            return
        f = self.send_head()
        if f:
            try:
                self.copyfile(f, self.wfile)
            except ConnectionResetError:
                pass
            f.close()

    def _download_new(self, path):
        # save file in local cache
        maxtries = 3
        head, _ = os.path.split(path)
        os.makedirs(head, exist_ok=True)
        totalsize = -1
        downloaded = 0
        for trynum in range(maxtries):
            try:
                headers = {}
                if downloaded > 0:
                    # if file was partly downloaded, only request the rest
                    headers["Range"] = f"bytes={downloaded}-"
                req = urllib.request.Request(
                    "http://snapshot.debian.org/" + self.path, headers=headers
                )
                # we use os.fdopen(os.open(...)) because we don't want to
                # truncate the file and seek to the right position but also
                # create it if it doesn't exist yet
                with urllib.request.urlopen(req) as f, os.fdopen(
                    os.open(path, os.O_RDWR | os.O_CREAT), "rb+"
                ) as out:
                    out.seek(downloaded)
                    if trynum == 0:
                        self.send_response(HTTPStatus.OK)
                        self.send_header("Content-type", f.headers["Content-type"])
                        self.send_header("Content-Length", f.headers["Content-Length"])
                        self.send_header("Last-Modified", f.headers["Last-Modified"])
                        self.end_headers()
                        totalsize = int(f.headers["Content-Length"])
                    while downloaded < totalsize:
                        chunksize = 200 * 1024
                        if totalsize - downloaded < chunksize:
                            chunksize = totalsize - downloaded
                        buf = f.read(chunksize)  # 200 kB/s
                        if len(buf) != chunksize:
                            # something went wrong
                            logging.warning(
                                "%s: wanted %d but got %d bytes (try %d of %d)",
                                path,
                                chunksize,
                                len(buf),
                                trynum + 1,
                                maxtries,
                            )
                            sleep(10)
                            break
                        sleep(1)  # snapshot.d.o needs heavy throttling
                        out.write(buf)
                        self.wfile.write(buf)
                        downloaded += chunksize
            except urllib.error.HTTPError as e:
                if e.code == 404:
                    self.send_error(404, "URLError")
                    return
                logging.warning("got urllib.error.HTTPError: %s %s", repr(e), self.path)
            except urllib.error.URLError as e:
                logging.warning("got urllib.error.URLError: %s", repr(e))
            if downloaded == totalsize:
                break
        if totalsize != downloaded:
            if os.path.exists(path):
                os.unlink(path)
            self.send_error(500, "URLError")
            return

    def log_message(self, fmt, *args):
        pass


def srcpkgversions_by_timestamp(srcpkgname, timestamp, suite):
    versions = set()
    timestamp_str = timestamp.strftime("%Y%m%dT%H%M%SZ")
    r = requests.get(
        f"http://snapshot.debian.org/archive/debian/{timestamp_str}"
        f"/dists/{suite}/main/source/Sources.xz",
        timeout=60,
    )
    data = lzma.decompress(r.content)
    for src in debian.deb822.Sources.iter_paragraphs(io.BytesIO(data)):
        if src["Package"] != srcpkgname:
            continue
        versions.add(debian.debian_support.Version(src["Version"]))
    return versions


def binpkgversion_by_timestamp(binpkgname, timestamp, suite, architecture):
    timestamp_str = timestamp.strftime("%Y%m%dT%H%M%SZ")
    r = requests.get(
        f"http://snapshot.debian.org/archive/debian/{timestamp_str}"
        f"/dists/{suite}/main/binary-{architecture}/Packages.xz",
        timeout=60,
    )
    data = lzma.decompress(r.content)
    for pkg in debian.deb822.Packages.iter_paragraphs(io.BytesIO(data)):
        if pkg["Package"] == binpkgname:
            return debian.debian_support.Version(pkg["Version"])
    return None


# This function does something similar to what this wiki page describes
# https://wiki.debian.org/BisectDebian#Finding_dates_for_specific_packages
#
# The problem with the approach on that wiki page as well as the one below in
# Python is, that it relies on the first_seen entry provided by snapshot.d.o.
# This means that we do not know when a package first appeared in a given
# suite. It could've first appeared in experimental or even in Debian Ports.
#
# Also see: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=806329
def first_seen_by_pkg(packages, timestamp_begin, timestamp_end, suite, architecture):
    timestamps = set()
    for pkg in packages:
        logging.info("obtaining versions for %s", pkg)
        if pkg.startswith("src:"):
            pkg = pkg[4:]
            oldest_versions = srcpkgversions_by_timestamp(pkg, timestamp_begin, suite)
            if len(oldest_versions) == 0:
                logging.error(
                    "source package %s cannot be found in good timestamp", pkg
                )
                sys.exit(1)
            elif len(oldest_versions) == 1:
                oldest_version = oldest_versions.pop()
            else:
                oldest_version = min(oldest_versions)
            newest_versions = srcpkgversions_by_timestamp(pkg, timestamp_end, suite)
            if len(newest_versions) == 0:
                logging.error("source package %s cannot be found in bad timestamp", pkg)
                sys.exit(1)
            elif len(newest_versions) == 1:
                newest_version = newest_versions.pop()
            else:
                newest_version = max(newest_versions)

            for result in requests.get(
                f"http://snapshot.debian.org/mr/package/{pkg}/", timeout=60
            ).json()["result"]:
                if debian.debian_support.Version(result["version"]) < oldest_version:
                    continue
                if debian.debian_support.Version(result["version"]) > newest_version:
                    continue
                r = requests.get(
                    f"http://snapshot.debian.org/mr/package/{pkg}"
                    f"/{result['version']}/allfiles?fileinfo=1",
                    timeout=60,
                )
                logging.info("retrieving for: %s", result["version"])
                for fileinfo in [
                    fileinfo
                    for fileinfos in r.json()["fileinfo"].values()
                    for fileinfo in fileinfos
                ]:
                    if fileinfo["archive_name"] != "debian":
                        continue
                    timestamps.add(
                        datetime.strptime(fileinfo["first_seen"], "%Y%m%dT%H%M%S%z")
                    )
        else:
            oldest_version = binpkgversion_by_timestamp(
                pkg, timestamp_begin, suite, architecture
            )
            if oldest_version is None:
                logging.error(
                    "binary package %s cannot be found in good timestamp", pkg
                )
                sys.exit(1)
            newest_version = binpkgversion_by_timestamp(
                pkg, timestamp_end, suite, architecture
            )
            if newest_version is None:
                logging.error("binary package %s cannot be found in bad timestamp", pkg)
                sys.exit(1)
            r = requests.get(f"http://snapshot.debian.org/mr/binary/{pkg}/", timeout=60)
            for result in r.json()["result"]:
                if debian.debian_support.Version(result["version"]) < oldest_version:
                    continue
                if debian.debian_support.Version(result["version"]) > newest_version:
                    continue
                r = requests.get(
                    f"http://snapshot.debian.org/mr/binary/{pkg}"
                    f"/{result['version']}/binfiles?fileinfo=1",
                    timeout=60,
                )
                logging.info("retrieving for: %s", result["version"])
                hashes = [
                    e["hash"]
                    for e in r.json()["result"]
                    if e["architecture"] == architecture
                ]
                for fileinfo in [
                    fileinfo for h in hashes for fileinfo in r.json()["fileinfo"][h]
                ]:
                    if fileinfo["archive_name"] != "debian":
                        continue
                    timestamps.add(
                        datetime.strptime(fileinfo["first_seen"], "%Y%m%dT%H%M%S%z")
                    )
    return timestamps


def get_mirror(port, timestamp):
    timestamp_str = timestamp.strftime("%Y%m%dT%H%M%SZ")
    if port is not None:
        return f"http://127.0.0.1:{port}/archive/debian/{timestamp_str}"
    return f"http://snapshot.debian.org/archive/debian/{timestamp_str}"


def runtest_cmd(cmd, env):
    ret = 0
    output = b""
    try:
        # we only use the more complex Popen method if live output is required
        # for logging levels of INFO or lower
        if logging.root.isEnabledFor(logging.INFO):
            parent_fd, child_fd = pty.openpty()
            with subprocess.Popen(
                cmd,
                stdin=child_fd,
                stderr=child_fd,
                stdout=child_fd,
                close_fds=True,
                env=env,
            ) as process:
                buf = io.BytesIO()
                os.close(child_fd)
                while process.poll() is None:
                    ready, _, _ = select.select([parent_fd], [], [], 1)
                    if parent_fd in ready:
                        try:
                            data = os.read(parent_fd, 10240)
                        except OSError:
                            break
                        if not data:
                            break  # EOF
                        os.write(sys.stdout.fileno(), data)
                        buf.write(data)
                os.close(parent_fd)
                ret = process.wait()
                output = buf.getvalue()
        else:
            output = subprocess.check_output(cmd, stderr=subprocess.STDOUT, env=env)
    except subprocess.CalledProcessError as e:
        ret = e.returncode
        output = e.output
    return (ret, output)


def runtest(timestamp, staticargs, toupgrade=None, badtimestamp=None):
    goodmirror = get_mirror(staticargs.port, timestamp)
    env = {k: v for k, v in os.environ.items() if k.startswith("DEBIAN_BISECT_")}
    env["DEBIAN_BISECT_EPOCH"] = str(int(timestamp.timestamp()))
    env["DEBIAN_BISECT_TIMESTAMP"] = timestamp.strftime("%Y%m%dT%H%M%SZ")
    env["PATH"] = os.environ.get("PATH", "/usr/sbin:/usr/bin:/sbin:/bin")
    if staticargs.port is not None:
        env["DEBIAN_BISECT_MIRROR"] = goodmirror
    if staticargs.depends or staticargs.qemu:
        scriptname = "run_bisect"
        if staticargs.qemu:
            scriptname = "run_bisect_qemu"
        # first try run_bisect.sh from the directory where debbisect lives in
        # case we run this from a git clone
        run_bisect = os.path.join(
            os.path.dirname(os.path.realpath(__file__)), scriptname + ".sh"
        )
        if not os.path.exists(run_bisect):
            run_bisect = os.path.join("/usr/share/devscripts/scripts/", scriptname)
        cmd = [run_bisect]
        if staticargs.depends is not None:
            cmd.append(staticargs.depends)
        else:
            cmd.append("")
        cmd.extend(
            [
                staticargs.script,
                goodmirror,
                staticargs.architecture,
                staticargs.suite,
                staticargs.components,
            ]
        )
        if staticargs.qemu:
            cmd.extend([staticargs.qemu["memsize"], staticargs.qemu["disksize"]])
        if toupgrade:
            cmd.extend([get_mirror(staticargs.port, badtimestamp), toupgrade])
    else:
        # execute it directly if it's an executable file or if it there are no
        # shell metacharacters
        if (
            os.access(staticargs.script, os.X_OK)
            or re.search(r"[^\w@\%+=:,.\/-]", staticargs.script, re.ASCII) is None
        ):
            cmd = [staticargs.script]
        else:
            cmd = ["sh", "-c", staticargs.script]
    return runtest_cmd(cmd, env)


def get_log_fname(timestamp, goodbad, toupgrade=None):
    if toupgrade is None:
        return f"debbisect.{timestamp.strftime('%Y%m%dT%H%M%SZ')}.log.{goodbad}"
    return f"debbisect.{timestamp.strftime('%Y%m%dT%H%M%SZ')}.{toupgrade}.log.{goodbad}"


def write_log_symlink(goodbad, output, timestamp, toupgrade=None):
    fname = get_log_fname(timestamp, goodbad, toupgrade)
    with open(fname, "wb") as f:
        f.write(output)
    if goodbad == "good":
        if os.path.lexists("debbisect.log.good"):
            os.unlink("debbisect.log.good")
        os.symlink(fname, "debbisect.log.good")
    elif goodbad == "bad":
        if os.path.lexists("debbisect.log.bad"):
            os.unlink("debbisect.log.bad")
        os.symlink(fname, "debbisect.log.bad")


def bisect(good, bad, staticargs):
    # no idea how to split this function into parts without making it
    # unreadable
    # pylint: disable=too-many-statements
    diff = bad - good
    print(f"snapshot timestamp difference: {diff / timedelta(days=1)} days")

    stepnum = 1
    starttime = datetime.now(timezone.utc)

    steps = round(
        (math.log(diff.total_seconds()) - math.log(DINSTALLRATE)) / math.log(2) + 2
    )
    print(f"approximately {steps} steps left to test")
    # verify that the good timestamp is really good and the bad timestamp is really bad
    # we try the bad timestamp first to make sure that the problem exists
    if not staticargs.ignore_cached_results and os.path.exists(
        get_log_fname(good, "good")
    ):
        print(f"#{stepnum}: using cached results from {get_log_fname(good, 'good')}")
    else:
        print(f"#{stepnum}: trying known good {format_timestamp(good)}...")
        ret, output = runtest(good, staticargs)
        if ret != 0:
            write_log_symlink("bad", output, good)
            print(
                "good timestamp was actually bad -- see debbisect.log.bad for details"
            )
            return None
        write_log_symlink("good", output, good)
    stepnum += 1
    steps = round(
        (math.log(diff.total_seconds()) - math.log(DINSTALLRATE)) / math.log(2) + 1
    )
    timeleft = steps * (datetime.now(timezone.utc) - starttime) / (stepnum - 1)
    print(f"computation time left: {timeleft}")
    print(f"approximately {steps} steps left to test")
    if not staticargs.ignore_cached_results and os.path.exists(
        get_log_fname(bad, "bad")
    ):
        print(f"#{stepnum}: using cached results from {get_log_fname(bad, 'bad')}")
    else:
        print(f"#{stepnum}: trying known bad {format_timestamp(bad)}...")
        ret, output = runtest(bad, staticargs)
        if ret == 0:
            write_log_symlink("good", output, bad)
            print(
                "bad timestamp was actually good -- see debbisect.log.good for details"
            )
            return None
        write_log_symlink("bad", output, bad)
    stepnum += 1

    while True:
        diff = bad - good
        # One may be tempted to try and optimize this step by finding all the
        # packages that differ between the two timestamps and then finding
        # all the snapshot timestamps where the involved packages changed
        # in their version. But since dependencies can arbitrarily change
        # between two given timestamps, drawing in more packages or requiring
        # less packages, the only reliable method is really to strictly bisect
        # by taking the timestamp exactly between the two and not involve
        # other guessing magic.
        newts = sanitize_timestamp(good + diff / 2)
        if newts in [good, bad]:
            # If the middle timestamp mapped onto good or bad, then the
            # timestamps are very close to each other. Test if there is maybe
            # not another one between them by sanitizing the timestamp one
            # second before the bad one
            newts = sanitize_timestamp(bad - timedelta(seconds=1))
            if newts == good:
                break
        print(f"snapshot timestamp difference: {diff / timedelta(days=1)} days")
        steps = round(
            (math.log(diff.total_seconds()) - math.log(DINSTALLRATE)) / math.log(2) + 0
        )
        timeleft = steps * (datetime.now(timezone.utc) - starttime) / (stepnum - 1)
        print(f"computation time left: {timeleft}")
        print(f"approximately {steps} steps left to test")
        if not staticargs.ignore_cached_results and os.path.exists(
            get_log_fname(newts, "good")
        ):
            print(
                f"#{stepnum}: using cached result (was good)"
                f" from {get_log_fname(newts, 'good')}"
            )
            good = newts
        elif not staticargs.ignore_cached_results and os.path.exists(
            get_log_fname(newts, "bad")
        ):
            print(
                f"#{stepnum}: using cached result (was bad)"
                f" from {get_log_fname(newts, 'bad')}"
            )
            bad = newts
        else:
            print(f"#{stepnum}: trying {format_timestamp(newts)}...")
            ret, output = runtest(newts, staticargs)
            if ret == 0:
                print("test script output: good")
                write_log_symlink("good", output, newts)
                good = newts
            else:
                print("test script output: bad")
                write_log_symlink("bad", output, newts)
                bad = newts
        stepnum += 1
    return good, bad


def datetimestr(val):
    if val == "now":
        return datetime.now(timezone.utc)
    if val == "today":
        return datetime.combine(date.today(), time(0, 0, 0, 0), timezone.utc)

    # since py3 we don't need pytz to figure out the local timezone
    localtz = datetime.now(timezone.utc).astimezone().tzinfo

    # first trying known formats
    for fmt in [
        "%Y%m%dT%H%M%SZ",  # snapshot.debian.org style
        # ISO 8601
        "%Y-%m-%d",
        "%Y-%m-%dT%H:%M",
        "%Y-%m-%dT%H:%M:%S",
        "%Y-%m-%dT%H:%M:%S%z",
    ]:
        try:
            dt = datetime.strptime(val, fmt)
        except ValueError:
            continue
        else:
            # if we don't know the timezone, assume the local one
            if dt.tzinfo is None:
                dt = dt.replace(tzinfo=localtz)
            return dt

    # try parsing the debian/changelog datetime format as specified by RFC 2822
    # we cannot use strptime() because that honors the locale and RFC
    # 2822 requires that day and month names be the English abbreviations.
    try:
        dt = email.utils.parsedate_to_datetime(val)
    except TypeError:
        pass
    else:
        return dt

    # next, try parsing using dateutil.parser
    if HAVE_DATEUTIL:
        try:
            dt = dateutil.parser.parse(val)
        except ValueError:
            pass
        else:
            # if we don't know the timezone, assume the local one
            if dt.tzinfo is None:
                dt = dt.replace(tzinfo=localtz)
            return dt

    # if that didn't work, try freeform formats
    if HAVE_PARSEDATETIME:
        cal = parsedatetime.Calendar()
        dt, ret = cal.parseDT(val)
        if ret != 0:
            # if we don't know the timezone, assume the local one
            if dt.tzinfo is None:
                dt = dt.replace(tzinfo=localtz)
            return dt

    if not HAVE_DATEUTIL:
        logging.info("parsing date failed -- maybe install python3-dateutil")
    if not HAVE_PARSEDATETIME:
        logging.info("parsing date failed -- maybe install python3-parsedatetime")

    # all failed, we cannot parse this
    raise argparse.ArgumentTypeError(f"not a valid datetime: {val}")


def qemuarg(val):
    defaultmem = "1G"
    defaultdisk = "4G"
    ret = {"memsize": defaultmem, "disksize": defaultdisk}
    for keyval in val.split(","):
        # we use startswith() so that "defaults" can also be abbreviated (even
        # down to the empty string)
        if "defaults".startswith(keyval):
            ret["memsize"] = defaultmem
            ret["disksize"] = defaultdisk
            continue
        try:
            key, val = keyval.split("=", maxsplit=1)
        except ValueError as e:
            raise argparse.ArgumentTypeError(f"no key=val pair: {keyval}") from e
        if key not in ["memsize", "disksize"]:
            raise argparse.ArgumentTypeError(f"unknown key: {key}")
        if not re.fullmatch(r"\d+((k|K|M|G|T|P|E|Z|Y)(iB|B)?)?", val):
            raise argparse.ArgumentTypeError(f"cannot parse size value: {val}")
        ret[key] = val
    return ret


def scriptarg(val):
    if os.path.exists(val) and not os.access(val, os.X_OK):
        logging.warning("script %s is a file but not executable", val)
    return val


def read_pkglist(infile):
    result = {}
    with open(infile, encoding="utf8") as f:
        for line in f:
            pkg, version = line.split("\t")
            result[pkg] = version.strip()
    return result


def upgrade_single_package(toupgrade, goodpkgs, badpkgs, good, bad, staticargs):
    if toupgrade in goodpkgs:
        print(
            f"test upgrading {toupgrade} {goodpkgs[toupgrade]}"
            f" -> {badpkgs[toupgrade]}..."
        )
    else:
        print(f"test installing {toupgrade} {badpkgs[toupgrade]}...")
    newbadpkgpath = f"./debbisect.{good.strftime('%Y%m%dT%H%M%SZ')}.{toupgrade}.pkglist"
    if (
        not staticargs.ignore_cached_results
        and os.path.exists(newbadpkgpath)
        and os.path.exists(get_log_fname(good, "good", toupgrade))
    ):
        print(
            f"using cached result (was good)"
            f" from {get_log_fname(good, 'good', toupgrade)}"
        )
        if toupgrade in goodpkgs:
            print(f"  upgrading {toupgrade} does not cause the problem")
        else:
            print(f"  installing {toupgrade} does not cause the problem")
        return
    if (
        not staticargs.ignore_cached_results
        and os.path.exists(newbadpkgpath)
        and os.path.exists(get_log_fname(good, "bad", toupgrade))
    ):
        print(
            f"using cached result (was bad)"
            f" from {get_log_fname(good, 'bad', toupgrade)}"
        )
        print(f"  upgrading {toupgrade} triggered the problem")
    else:
        ret, output = runtest(good, staticargs, toupgrade, bad)
        if ret == 0:
            write_log_symlink("good", output, good, toupgrade)
            if toupgrade in goodpkgs:
                print(f"  upgrading {toupgrade} does not cause the problem")
            else:
                print(f"  installing {toupgrade} does not cause the problem")
            return
        write_log_symlink("bad", output, good, toupgrade)
        print(f"  upgrading {toupgrade} triggered the problem")
    # this package introduced the regression check if more than
    # just the package in question got upgraded
    newbadpkgs = read_pkglist(newbadpkgpath)
    # find all packages that are completely new or of a
    # different version than those in the last good test
    newupgraded = []
    for pkg, version in newbadpkgs.items():
        if pkg not in goodpkgs or version != goodpkgs[pkg]:
            newupgraded.append(pkg)
    if not newupgraded:
        logging.error("no difference -- this should never happen")
        sys.exit(1)
    elif len(newupgraded) == 1:
        # the only upgraded package should be the one that was
        # requested to be upgraded
        assert newupgraded[0] == toupgrade
    else:
        print("  additional packages that got upgraded/installed at the same time:")
        for newtoupgrade in newupgraded:
            if newtoupgrade == toupgrade:
                continue
            print(
                f"    {newtoupgrade} {goodpkgs.get(newtoupgrade, '(n.a.)')}"
                f" -> {newbadpkgs[newtoupgrade]}"
            )


def parseargs():
    progname = sys.argv[0]
    parser = argparse.ArgumentParser(
        formatter_class=argparse.RawDescriptionHelpFormatter,
        description=f"""\

Execute a script or a shell snippet for a known good timestamp and a known bad
timestamp and then bisect the timestamps until a timestamp from
snapshot.debian.org is found where the script first fails. Environment
variables are used to tell the script which timestamp to test. See ENVIRONMENT
VARIABLES below. At the end of the execution, the files debbisect.log.good and
debbisect.log.bad are the log files of the last good and last bad run,
respectively. By default, a temporary caching mirror is executed to reduce
bandwidth usage on snapshot.debian.org.  If you plan to run debbisect multiple
times on a similar range of timestamps, consider setting a non-temporary cache
directory with the --cache option.

The program has three basic modes of operation. In the first, the given script
is responsible to set up everything as needed:

    $ {progname} "last week" today script.sh
    $ diff -u debbisect.log.good debbisect.log.bad

If also the --depends option is given, then a chroot of the correct timestamp
will be created each time and the script will receive as first argument the
path to that chroot. Additionally, this mode allows debbisect to figure out the
exact package that was responsible for the failure instead of only presenting
you the last good and first bad timestamp.

Lastly, you can also provide the --qemu option. In this mode, your test will be
create a qemu virtual machine of the correct timestamp each time. The script
will receive the correct ssh config to log into a host named qemu and execute
arbitrary commands.

""",
        epilog=f"""\

*EXAMPLES*

Just run "do_something" which runs the test and returns a non-zero exit on
failure.

    $ {progname} "last week" today "mmdebstrap --customize-hook\
='chroot \\"\\$1\\" do_something' unstable - \\$DEBIAN_BISECT_MIRROR >/dev/null"
    $ diff -u debbisect.log.good debbisect.log.bad

Since the command can easily become very long and quoting very involved, lets
instead use a script:

    $ cat << END > script.sh
    > #!/bin/sh
    > set -exu
    > mmdebstrap \\
    > --verbose \\
    > --aptopt='Acquire::Check-Valid-Until "false"' \\
    > --variant=apt \\
    > --include=pkga,pkgb,pkgc \\
    > --customize-hook='chroot "$1" dpkg -l' \\
    > --customize-hook='chroot "$1" do_something' \\
    > unstable \\
    > - \\
    > $DEBIAN_BISECT_MIRROR \\
    > >/dev/null
    > END
    $ chmod +x script.sh
    $ {progname} --verbose --cache=./cache "two years ago" yesterday ./script.sh
    $ diff -u debbisect.log.good debbisect.log.bad
    $ rm -r ./cache

This example sets Acquire::Check-Valid-Until to not fail on snapshot timestamps
from "two years ago", uses the "apt" variant (only Essential:yes plus apt),
installs the packages required for the test using --include, runs "dpkg -l" so
that we can see which packages differed in the logs at the end and uses
--cache=cache so that the apt cache does not get discarded at the end and the
command can be re-run without downloading everything from snapshot.debian.org
again.

If you want to build a source package you can use the script shipped by
devscripts as /usr/share/doc/devscripts/examples/debbisect_buildsrc.sh and
either use it unmodified like this:

    $ DEBIAN_BISECT_SRCPKG=mysrc {progname} "two years ago" yesterday \
    > /usr/share/doc/devscripts/examples/debbisect_buildsrc.sh

or use the script as a starting point to do your own custom builds.

Once debbisect has finished bisecting and figured out the last good and the
first bad timestamp, there might be more than one package that differs in
version between these two timestamps. debbisect can figure out which package is
the culprit if you hand it control over installing dependencies for you via the
--depends option. With that option active, the script will not be responsible
to set up a chroot itself but is given the path to an existing chroot as the
first argument. Here is a real example that verifies the package responsible
for Debian bug #912935:

    $ {progname} --depends=botch "2018-11-17" "2018-11-22" \
'chroot "$1" botch-dose2html --packages=/dev/null --help'
    [...]
    test upgrading python3-minimal 3.6.7-1 -> 3.7.1-2...
      upgrading python3-minimal triggered the problem

If you want to run above test under qemu, then you would run:

    $ {progname} --depends=botch --qemu=defaults "2018-11-17" "2018-11-22" \
'ssh -F "$1" qemu botch-dose2html --packages=/dev/null --help'

In the last two examples we omitted the --cache argument for brevity. But
please make use of it to reduce the load on snapshot.debian.org.

*TIMESTAMPS*

Valid good and bad timestamp formats are either:

    > the format used by snapshot.debian.org
    > ISO 8601 (with or without time, seconds and timezone)
    > RFC 2822 (used in debian/changelog)
    > all formats understood by the Python dateutil module (if installed)
    > all formats understood by the Python parsedatetime module (if installed)

Without specifying the timezone explicitly, the local offset is used.

Examples (corresponding to the items in above list, respectively):

    > 20200313T065326Z
    > 2020-03-13T06:53:26+00:00
    > Fri, 29 Nov 2019 14:00:08 +0100
    > 5:50 A.M. on June 13, 1990
    > two weeks ago

The earliest timestamp that works with debbisect should be 2006-08-10.

*ENVIRONMENT VARIABLES*

The following environment variables are available to the test script:

DEBIAN_BISECT_MIRROR    Contains the caching mirror address.

DEBIAN_BISECT_EPOCH     Contains an integer representing the unix epoch of the
                        current timestamp. The value of this variable can
                        directly be assigned to SOURCE_DATE_EPOCH.

DEBIAN_BISECT_TIMESTAMP Contains a timestamp in the format used by
                        snapshot.debian.org. Can also be generated from
                        DEBIAN_BISECT_EPOCH via:
                        date --utc --date=@$DEBIAN_BISECT_EPOCH +%Y%m%dT%H%M%SZ

DEBIAN_BISECT_*         All environment variables starting with DEBIAN_BISECT_
                        are passed to the test script.

Written by Johannes Schauer Marin Rodrigues <josch@debian.org>
""",
    )
    parser.add_argument(
        "-d",
        "--debug",
        help="Print lots of debugging statements",
        action="store_const",
        dest="loglevel",
        const=logging.DEBUG,
        default=logging.WARNING,
    )
    parser.add_argument(
        "-v",
        "--verbose",
        help="Be verbose",
        action="store_const",
        dest="loglevel",
        const=logging.INFO,
    )
    parser.add_argument(
        "--cache", help="cache directory -- by default $TMPDIR is used", type=str
    )
    parser.add_argument("--nocache", help="disable cache", action="store_true")
    parser.add_argument(
        "--port",
        help="manually choose port number for the apt cache instead of "
        "automatically choosing a free port",
        type=int,
        default=0,
    )
    parser.add_argument(
        "--depends",
        help="Comma separated list of binary packages the test script "
        "requires. With this option, the test script will run inside a "
        "chroot with the requested packages installed.",
        type=str,
    )
    parser.add_argument(
        "--qemu",
        help="Create qemu virtual machine and pass a ssh config file to the "
        "test script. This argument takes a comma-separated series of "
        "key=value pairs to specify the virtual machine memory size (via "
        "memsize) and the virtual machine disksize (via disksize). Sizes "
        "are measured in bytes or with common unit suffixes like M or G. "
        "To pick the default values (disksize=4G,memsize=1G) the shorthand "
        "'defaults' can be passed.",
        type=qemuarg,
    )
    parser.add_argument(
        "--architecture",
        help="Chosen architecture when creating the chroot with --depends or "
        "--qemu (default: native architecture)",
        default=subprocess.check_output(["dpkg", "--print-architecture"]).rstrip(),
        type=str,
    )
    parser.add_argument(
        "--suite",
        help="Chosen suite when creating the chroot with --depends or --qemu "
        "(default: unstable)",
        default="unstable",
        type=str,
    )
    parser.add_argument(
        "--components",
        help="Chosen components (separated by comma or whitespace) when "
        "creating the chroot with --depends or --qemu (default: main)",
        default="main",
        type=str,
    )
    parser.add_argument(
        "--no-find-exact-package",
        help="Normally, when the --depends argument is given so that "
        "debbisect takes care of managing dependencies, the precise package "
        "that introduced the problem is determined after bisection by "
        "installing the packages that differ between the last good and "
        "first bad timestamp one by one. This option disables this feature.",
        action="store_true",
    )
    parser.add_argument(
        "--ignore-cached-results",
        help="Perform a run for a timestamp even if a log file for it exists "
        "in the current directory",
        action="store_true",
    )
    parser.add_argument(
        "good",
        type=datetimestr,
        help="good timestamp -- see section TIMESTAMPS for valid formats",
    )
    parser.add_argument(
        "bad",
        type=datetimestr,
        help="bad timestamp -- see section TIMESTAMPS for valid formats",
    )
    parser.add_argument(
        "script",
        type=scriptarg,
        help="test script -- can either be a shell code snippet or an "
        "executable script. A non zero exit code indicates failure. "
        "When also --depends is used, then the first argument to the "
        "script will be the chroot directory. When --qemu is used, then "
        "the first argument to the script will an ssh config for a host "
        "named qemu.",
    )
    return parser.parse_args()


def setupcache(cache, port):
    if cache:
        cachedir = cache
    else:
        cachedir = tempfile.mkdtemp(prefix="debbisect")
    logging.info("using cache directory: %s", cachedir)
    os.makedirs(cachedir + "/pool", exist_ok=True)
    httpd = socketserver.TCPServer(
        # the default address family for socketserver is AF_INET so we
        # explicitly bind to ipv4 localhost
        ("127.0.0.1", port),
        partial(Proxy, directory=cachedir),
        # to avoid "Address already in use" when the port is specified
        # manually, we set socket.SO_REUSEADDR
        # to do so, we must set allow_reuse_address and then bind and
        # activate manually
        bind_and_activate=False,
    )
    # this sets socket.SO_REUSEADDR
    httpd.allow_reuse_address = True
    httpd.server_bind()
    httpd.server_activate()
    # run server in a new thread
    server_thread = threading.Thread(target=httpd.serve_forever)
    server_thread.daemon = True
    # start thread
    server_thread.start()
    # retrieve port (in case it was generated automatically)
    _, port = httpd.server_address

    def teardown():
        httpd.shutdown()
        httpd.server_close()
        server_thread.join()
        if not cache:
            # this should be a temporary directory but lets still be super
            # careful
            if os.path.exists(cachedir + "/pool"):
                shutil.rmtree(cachedir + "/pool")
            if os.path.exists(cachedir + "/archive"):
                shutil.rmtree(cachedir + "/archive")
            os.rmdir(cachedir)

    return port, teardown


def find_exact_package(good, bad, staticargs, depends, no_find_exact_package):
    goodpkglist = f"./debbisect.{good.strftime('%Y%m%dT%H%M%SZ')}.pkglist"
    if not os.path.exists(goodpkglist):
        logging.error("%s doesn't exist", goodpkglist)
        sys.exit(1)
    badpkglist = f"./debbisect.{bad.strftime('%Y%m%dT%H%M%SZ')}.pkglist"
    if not os.path.exists(badpkglist):
        logging.error("%s doesn't exist", badpkglist)
        sys.exit(1)

    # Create a set of packages for which either the version in the last good
    # and first bad run differs or which only exist in the first bad run.
    goodpkgs = read_pkglist(goodpkglist)
    badpkgs = read_pkglist(badpkglist)
    upgraded = set()
    for pkg, version in goodpkgs.items():
        if pkg in badpkgs and version != badpkgs[pkg]:
            upgraded.add(pkg)
    for pkg, version in badpkgs.items():
        if pkg not in goodpkgs or version != goodpkgs[pkg]:
            upgraded.add(pkg)
    upgraded = list(sorted(upgraded))
    if not upgraded:
        logging.error("no difference -- this should never happen")
        sys.exit(1)
    elif len(upgraded) == 1:
        print(
            f"only one package differs: {upgraded[0]}"
            f" {goodpkgs.get(upgraded[0], '(n.a.)')} -> {badpkgs[upgraded[0]]}"
        )
    else:
        print(
            "the following packages differ between the last good and "
            "first bad timestamp:"
        )
        for toupgrade in upgraded:
            print(
                f"  {toupgrade} {goodpkgs.get(toupgrade, '(n.a.)')}"
                f" -> {badpkgs[toupgrade]}"
            )

        # if debbisect was tasked with handling dependencies itself, try to
        # figure out the exact package that introduce the break
        if depends and not no_find_exact_package:
            for toupgrade in upgraded:
                upgrade_single_package(
                    toupgrade, goodpkgs, badpkgs, good, bad, staticargs
                )


def ensure_mmdebstrap_version(reqver: str) -> bool:
    try:
        version = subprocess.check_output(
            ["mmdebstrap", "--version"], stderr=subprocess.DEVNULL
        )
    except subprocess.CalledProcessError:
        print("running mmdebstrap --version resulted in non-zero exit")
        sys.exit(1)
    except FileNotFoundError:
        print("you need to install mmdebstrap")
        sys.exit(1)

    version = version.decode().removeprefix("mmdebstrap ")

    return debian.debian_support.Version(version) >= debian.debian_support.Version(
        reqver
    )


def main():
    args = parseargs()

    logging.basicConfig(level=args.loglevel)

    good = sanitize_timestamp(args.good)
    if good != args.good:
        print(
            f"good timestamp {format_timestamp(args.good)} was remapped to"
            f" snapshot.d.o timestamp {format_timestamp(good)}"
        )
    bad = sanitize_timestamp(args.bad)
    if bad != args.bad:
        print(
            f"bad timestamp {format_timestamp(args.bad)} was remapped to"
            f" snapshot.d.o timestamp {format_timestamp(bad)}"
        )

    if good > bad:
        print("good is later than bad")
        sys.exit(1)

    # check if mmdebstrap is installed and at least 1.3.0
    if (args.depends or args.qemu) and not ensure_mmdebstrap_version("1.3.0"):
        print("you need at least mmdebstrap version 1.3.0")
        sys.exit(1)

    port = None
    if not args.nocache:
        port, teardown = setupcache(args.cache, args.port)
        atexit.register(teardown)

    staticargs = collections.namedtuple(
        "args",
        [
            "script",
            "port",
            "depends",
            "architecture",
            "suite",
            "components",
            "qemu",
            "ignore_cached_results",
        ],
    )
    for a in staticargs._fields:
        setattr(staticargs, a, getattr(args, a))
    staticargs.port = port
    if good == bad:
        # test only single timestamp
        print(f"trying single timestamp {format_timestamp(good)}...")
        if not staticargs.ignore_cached_results and os.path.exists(
            get_log_fname(good, "good")
        ):
            print(f"using cached result (was good) from {get_log_fname(good, 'good')}")
            ret = 0
        elif not staticargs.ignore_cached_results and os.path.exists(
            get_log_fname(good, "bad")
        ):
            print(f"using cached result (was bad) from {get_log_fname(good, 'bad')}")
            ret = 1
        else:
            ret, output = runtest(good, staticargs)
            if ret == 0:
                print("test script output: good")
                write_log_symlink("good", output, good)
            else:
                print("test script output: bad")
                write_log_symlink("bad", output, good)
        sys.exit(ret)
    res = bisect(good, bad, staticargs)
    if res is not None:
        good, bad = res
        print("bisection finished successfully")
        print(f"  last good timestamp: {format_timestamp(good)}")
        print(f"  first bad timestamp: {format_timestamp(bad)}")

        find_exact_package(
            good, bad, staticargs, args.depends, args.no_find_exact_package
        )


if __name__ == "__main__":
    main()